Last week you said we are missing a factor of 1000.
Now you say are missing a factor of 50!!!
Is there another factor we are not considering? Why weren’t you able to get your FPGA working? What FPGA were you using?
Thanks for the explanation. I should have imagined that the huge amount of RAM cannot possibly fit into a chip. I’ve never seen one ASIC chip, so easily misled by ChatGPT. But if this GrinderPro is really someone in Shenzhen University, they could easily invite me to verify their claim. It may just take me an hour to get to their place, depending which which campus they are in.
How will you verify their claim? Even if the FPGA unit works they could have expended a huge amount of energy preplotting. This all seems like an elaborate scam except for the fact they are not asking for any money for the USA license. Not sure what to make of it.
This might be what they mean when they said they are not “simple” grinding.
The only hint they gave is they are using multiplexing. Could they be sharing the memory with the host system with 2TB of RAM and partitioning the FPGA RAM into ROM?
I obviously will not verify the energy part. But as long as they can generate a large amount of plots on the fly without hard drives and Internet connection, that is already very interesting to look at.
That’s still a lot… and it partly comes at the cost of increased pre-plotting time.
In theory you can get the 1000x, but the pre-plotting time would be near infinite.
When NoSSD 2.0 came out I stopped working on it. It’s just too much work…
Kria K26
I think the biggest you can get right now is the H200 with 141 GB of HBM, that’s kinda “on-chip”. The biggest HBM FPGA you can get is probably 64 GB.
I’m pretty sure the 2 TB RAM is just for finding plot ids that pass 3 signage points.
Maybe, but then it’s still off by a factor of 50.
Not including the cost to pre-plot, which will take months of running the FPGAs in pre-plotting mode. And you need roughly 3-5 PiB to store the data for 300 PiB effective farm size.
Not for long. 3D stacked HBM is coming soon.
https://spectrum.ieee.org/hybrid-bonding
Yeah they are definitely not doing this. It is something else. Probably a bug in the plot verification. IDK
Given recent Chia price, and the longterm trend for it to keep dropping, this FPGA would have to be very low cost to ever break even in the announced CNI timeframe for the new plot formats.
And that’s BEFORE factoring in the electric cost to run it.
If the claimed efficiency is real, it will break even. But that’s a big if that we’re all wondering about here.
We not affiliated with NoSSD. GrinderPro does not have right from licensing our code. They only contractor. Men paid fair for time.
Could you try to affiliate yourselves with nossd?
They are the GOAT (greatest of all time)
Shenzhen is a core hub for IT industry in China. There’s much that can be done with all the resources available.
Thats complete nonsense, learn to comprehend written word!
1 Like
I was thinking about L1 cache for a reference since it’s pretty much the fastest memory access option with lowest latency. Since even at the maximum speed of light it would create a 1ns delay each way if a GPU/was 1ft away with no other obstructions like signal degradation. Meaning there’s absolutely no way they could go past that theoretical maximum. And even if they had reached the maximum then IBM/Intel/AMD/Nvidia would like to offer them more money then they could ever make off Chia.
The other thing with interconnects is that even the worlds (current) fastest super computer (Frontier) has 12.8Tib/s custom interconnects with maximum 75TB/s write and 35TB/s read to all drives. So if they’re connecting multiple FPGA’s/writing to storage and somehow are beating the Frontier Super Computer, they again would have way better money making opportunities that Chia.
Even with the upcoming HBM3e on chip memory that will be coming to Nvidia data center GPU’s in Q4 you are getting 1.2TB/s on a 1024pin package that’s 11mm x 11mm. so for the 72TiB you’d need 60 packages which would be minimum 8.5cm x 8.5cm of space just for HBM memory. And don’t forget you need cooling since it would be running at 100% all the time and drawing ~300+ watts just for the 60 memory packages alone. And if they were all squeezed together in 8.5x8.5cm they’d be touching and… well there’s a reason they’re spaced on GPU’s even though clustering all right next to the GPU’s would reduce trace length.
With L1 cache on a 3ghz processor if my calculations are correct you’d max out at 24GB/s per MB cache. That means you would need 3GB of l1 cache speed memory for 72TiB/s. And you’d have to figure out how to cram all the plot calculations into that space. Even Intels 144 Xeon 6780E is only built with 13.5MB L1 cache. AMD maxes out at 8MB L1 and 768MB of the much much slower L3 cache on their latest EPYC 128 core CPU and that’s using their 3D cache technology.
Even on a 3nm node you’d be at around 6.25MB l1 cache per square mm. So to reach 3GB l1 you’d need 491mm of die space just for l1 cache (2.2cmx2.2cm).
Could they have done these things. Again maybe yes with cutting edge tech that somehow hasn’t been mentioned anywhere before. Would they be able to find a fab with spare capacity that could make the actual FPGA… Extremely doubtful. But most importantly. All these advances would make them massively more money by licensing it to Intel/AMD/Nvidia or even by improving China CPU’s like Loogson so they’re not 4 years behind Intel (making smuggling CPU’s into China a real constant thing) rather than using it to farm Chia. Even at $100 a coin with 1 million coins they’d make so much more money selling this tech. And then they’d also not have to post in Chia groups looking for “joint investors” rather than rolling in cache (pun intended)
2 Likes
Exactly what I was thinking.
I’m flying into Shenzhen to take care of some business on September 1. I’m hoping to meet with the GrinderPro team later in the week.
I have been talking to them daily for the past week and I’m still not sure what they are doing. From what I can gather they are not really grinding the plots in the traditional sense.
3 Likes
Can you send a public key which would allow us to see your farm on the blockchain? pool_public_key, pool_contract_puzzle_hash, farmer_public_key, or similar? This would help validate some of these claims.
2 Likes
It’s a joke bud, lighten up. Anyone who doesn’t see the humor in picking up coinbros romantically isn’t any fun.
1 Like
@RaYvA The trouble these days is people read things into something its not, it was light hearted humour, a play on words, get over it.
2 Likes
I was not!
Was it light-hearted in 1943?